Raphael
commited on
Improve translation and subtitles sync
Browse filesSigned-off-by: Raphael <oOraph@users.noreply.github.com>
app.py
CHANGED
@@ -10,6 +10,7 @@ import gradio as gr
|
|
10 |
import moviepy.editor as mp
|
11 |
import numpy as np
|
12 |
import pysrt
|
|
|
13 |
import torch
|
14 |
from transformers import pipeline
|
15 |
import yt_dlp
|
@@ -22,9 +23,10 @@ logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(level
|
|
22 |
LOG = logging.getLogger(__name__)
|
23 |
CLIP_SECONDS = 20
|
24 |
SLICES = 4
|
25 |
-
SLICE_DURATION = CLIP_SECONDS / SLICES
|
26 |
# At most 6 mins
|
27 |
MAX_CHUNKS = 45
|
|
|
28 |
|
29 |
asr_kwargs = {
|
30 |
"task": "automatic-speech-recognition",
|
@@ -118,7 +120,7 @@ def process_video(basedir: str, duration, translate: bool):
|
|
118 |
subs = translation(transcriptions, translate)
|
119 |
srt_file = build_srt_clips(subs, basedir)
|
120 |
summary = summarize(transcriptions, translate)
|
121 |
-
return srt_file, ' '.join(subs).strip(), summary
|
122 |
|
123 |
|
124 |
def transcription(audio_dir: str, duration):
|
@@ -141,74 +143,131 @@ def transcription(audio_dir: str, duration):
|
|
141 |
t = asr(d, max_new_tokens=10000)
|
142 |
transcriptions.extend(t)
|
143 |
|
144 |
-
transcriptions = [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
145 |
elapsed = time.time() - start
|
146 |
LOG.info("Transcription done, elapsed %.2f seconds", elapsed)
|
147 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
148 |
|
149 |
|
150 |
def translation(transcriptions, translate):
|
|
|
151 |
if translate:
|
152 |
LOG.info("Performing translation")
|
153 |
start = time.time()
|
154 |
-
translations = translator(transcriptions)
|
155 |
-
|
|
|
|
|
|
|
156 |
elapsed = time.time() - start
|
157 |
LOG.info("Translation done, elapsed %.2f seconds", elapsed)
|
|
|
158 |
else:
|
159 |
-
|
160 |
-
return
|
161 |
|
162 |
|
163 |
def summarize(transcriptions, translate):
|
164 |
LOG.info("Generating video summary")
|
165 |
-
whole_text = ' '.join(
|
166 |
-
word_count = len(whole_text.split())
|
167 |
summary = summarizer(whole_text)
|
168 |
# min_length=word_count // 4 + 1,
|
169 |
# max_length=word_count // 2 + 1)
|
170 |
-
summary = translation([summary[0]['summary_text']], translate)[0]
|
171 |
-
return summary
|
172 |
|
173 |
|
174 |
-
def
|
175 |
-
LOG.info("Building srt segments")
|
176 |
-
|
177 |
for sub in subtitles:
|
178 |
-
chunks = np.array_split(sub.split(' '), SLICES)
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
'text': c.strip(),
|
190 |
-
'start': i * SLICE_DURATION,
|
191 |
-
'end': (i + 1) * SLICE_DURATION
|
192 |
-
})
|
193 |
-
|
194 |
-
return segments
|
195 |
|
196 |
|
197 |
-
def build_srt_clips(
|
198 |
|
199 |
LOG.info("Generating subtitles")
|
200 |
-
segments =
|
201 |
|
202 |
LOG.info("Building srt clips")
|
203 |
-
max_text_len =
|
204 |
subtitles = pysrt.SubRipFile()
|
205 |
-
first = True
|
206 |
for segment in segments:
|
207 |
-
start = segment['start']
|
208 |
-
|
209 |
-
start += 3000
|
210 |
-
first = False
|
211 |
-
end = segment['end'] * 1000
|
212 |
text = segment['text']
|
213 |
text = text.strip()
|
214 |
if len(text) < max_text_len:
|
@@ -250,4 +309,5 @@ iface = gr.Interface(
|
|
250 |
gr.Text(label="Full transcription")
|
251 |
])
|
252 |
|
|
|
253 |
iface.launch()
|
|
|
10 |
import moviepy.editor as mp
|
11 |
import numpy as np
|
12 |
import pysrt
|
13 |
+
import re
|
14 |
import torch
|
15 |
from transformers import pipeline
|
16 |
import yt_dlp
|
|
|
23 |
LOG = logging.getLogger(__name__)
|
24 |
CLIP_SECONDS = 20
|
25 |
SLICES = 4
|
26 |
+
# SLICE_DURATION = CLIP_SECONDS / SLICES
|
27 |
# At most 6 mins
|
28 |
MAX_CHUNKS = 45
|
29 |
+
SENTENCE_SPLIT = re.compile(r'([^.?!]*[.?!]+)([^.?!].*|$)')
|
30 |
|
31 |
asr_kwargs = {
|
32 |
"task": "automatic-speech-recognition",
|
|
|
120 |
subs = translation(transcriptions, translate)
|
121 |
srt_file = build_srt_clips(subs, basedir)
|
122 |
summary = summarize(transcriptions, translate)
|
123 |
+
return srt_file, ' '.join([s['text'].strip() for s in subs]).strip(), summary
|
124 |
|
125 |
|
126 |
def transcription(audio_dir: str, duration):
|
|
|
143 |
t = asr(d, max_new_tokens=10000)
|
144 |
transcriptions.extend(t)
|
145 |
|
146 |
+
transcriptions = [
|
147 |
+
{
|
148 |
+
'text': t['text'].strip(),
|
149 |
+
'start': i * CLIP_SECONDS * 1000,
|
150 |
+
'end': (i + 1) * CLIP_SECONDS * 1000
|
151 |
+
} for i, t in enumerate(transcriptions)
|
152 |
+
]
|
153 |
+
|
154 |
+
if transcriptions:
|
155 |
+
transcriptions[0]['start'] += 2500
|
156 |
+
|
157 |
+
# Will improve the translation
|
158 |
+
segments = segments_on_sentence_boundaries(transcriptions)
|
159 |
+
|
160 |
elapsed = time.time() - start
|
161 |
LOG.info("Transcription done, elapsed %.2f seconds", elapsed)
|
162 |
+
return segments
|
163 |
+
|
164 |
+
|
165 |
+
def segments_on_sentence_boundaries(segments):
|
166 |
+
|
167 |
+
LOG.info("Segmenting along sentence boundaries for better translations")
|
168 |
+
|
169 |
+
new_segments = []
|
170 |
+
i = 0
|
171 |
+
while i < len(segments):
|
172 |
+
s = segments[i]
|
173 |
+
text = s['text'].strip()
|
174 |
+
if not text:
|
175 |
+
i += 1
|
176 |
+
continue
|
177 |
+
|
178 |
+
if i == len(segments)-1:
|
179 |
+
new_segments.append(s)
|
180 |
+
break
|
181 |
+
|
182 |
+
next_s = segments[i+1]
|
183 |
+
|
184 |
+
next_text = next_s['text'].strip()
|
185 |
+
if not next_text or (text[-1] in ['.', '?', '!']):
|
186 |
+
new_segments.append(s)
|
187 |
+
i += 1
|
188 |
+
continue
|
189 |
+
|
190 |
+
m = SENTENCE_SPLIT.match(next_s['text'].strip())
|
191 |
+
if not m:
|
192 |
+
LOG.warning("Bad pattern matching on segment [%s], "
|
193 |
+
"this should not be possible", next_s['text'])
|
194 |
+
s['end'] = next_s['end']
|
195 |
+
s['text'] = '{} {}'.format(s['text'].strip(), next_s['text'].strip())
|
196 |
+
new_segments.append(s)
|
197 |
+
i += 2
|
198 |
+
else:
|
199 |
+
before = m.group(1)
|
200 |
+
after = m.group(2)
|
201 |
+
next_segment_duration = next_s['end'] - next_s['start']
|
202 |
+
ratio = len(before) / len(next_text)
|
203 |
+
add_time = int(next_segment_duration * ratio)
|
204 |
+
s['end'] = s['end'] + add_time
|
205 |
+
s['text'] = '{} {}'.format(text, before)
|
206 |
+
next_s['start'] = next_s['start'] + add_time
|
207 |
+
next_s['text'] = after.strip()
|
208 |
+
new_segments.append(s)
|
209 |
+
i += 1
|
210 |
+
|
211 |
+
return new_segments
|
212 |
|
213 |
|
214 |
def translation(transcriptions, translate):
|
215 |
+
translations_d = []
|
216 |
if translate:
|
217 |
LOG.info("Performing translation")
|
218 |
start = time.time()
|
219 |
+
translations = translator([t['text'] for t in transcriptions])
|
220 |
+
for i, t in enumerate(transcriptions):
|
221 |
+
tsl = t.copy()
|
222 |
+
tsl['text'] = translations[i]['translation_text'].strip()
|
223 |
+
translations_d.append(tsl)
|
224 |
elapsed = time.time() - start
|
225 |
LOG.info("Translation done, elapsed %.2f seconds", elapsed)
|
226 |
+
LOG.info('Translations %s', translations_d)
|
227 |
else:
|
228 |
+
translations_d = transcriptions
|
229 |
+
return translations_d
|
230 |
|
231 |
|
232 |
def summarize(transcriptions, translate):
|
233 |
LOG.info("Generating video summary")
|
234 |
+
whole_text = ' '.join([t['text'].strip() for t in transcriptions])
|
235 |
+
# word_count = len(whole_text.split())
|
236 |
summary = summarizer(whole_text)
|
237 |
# min_length=word_count // 4 + 1,
|
238 |
# max_length=word_count // 2 + 1)
|
239 |
+
summary = translation([{'text': summary[0]['summary_text']}], translate)[0]
|
240 |
+
return summary['text']
|
241 |
|
242 |
|
243 |
+
def segment_slices(subtitles: list[str]):
|
244 |
+
LOG.info("Building srt segments slices")
|
245 |
+
slices = []
|
246 |
for sub in subtitles:
|
247 |
+
chunks = np.array_split(sub['text'].split(' '), SLICES)
|
248 |
+
start = sub['start']
|
249 |
+
duration = sub['end'] - start
|
250 |
+
for i in range(0, SLICES):
|
251 |
+
s = {
|
252 |
+
'text': ' '.join(chunks[i]),
|
253 |
+
'start': start + i * duration / SLICES,
|
254 |
+
'end': start + (i+1) * duration / SLICES
|
255 |
+
}
|
256 |
+
slices.append(s)
|
257 |
+
return slices
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
|
259 |
|
260 |
+
def build_srt_clips(segments, basedir):
|
261 |
|
262 |
LOG.info("Generating subtitles")
|
263 |
+
segments = segment_slices(segments)
|
264 |
|
265 |
LOG.info("Building srt clips")
|
266 |
+
max_text_len = 45
|
267 |
subtitles = pysrt.SubRipFile()
|
|
|
268 |
for segment in segments:
|
269 |
+
start = segment['start']
|
270 |
+
end = segment['end']
|
|
|
|
|
|
|
271 |
text = segment['text']
|
272 |
text = text.strip()
|
273 |
if len(text) < max_text_len:
|
|
|
309 |
gr.Text(label="Full transcription")
|
310 |
])
|
311 |
|
312 |
+
# iface.launch(server_name="0.0.0.0", server_port=6443)
|
313 |
iface.launch()
|